How does GCC access global variables on x86

[Xin Tong]

Is there a performance penalty for accessing global with rip on
architecture as new as nehalem, in comparsion to access variables
using rsp?

>>>>   4006f9:       f0 80 0d bf 0b 20 00    lock orb $0x0,0x200bbf(%rip)
>>>>      # 6012c0 <t2lockor>
>>>>   400700:       00

>>>>   4006f9:       f0 80 0d bf 0b 20 00    lock orb $0x0,0x0(%rsp)
>>>>      # 6012c0 <t2lockor>
>>>>   400700:       00

The later seems to perform better.

Thanks

Xin


On Tue, Feb 7, 2012 at 12:44 PM, Andrew Haley <aph@redhat.com> wrote:
> On 02/07/2012 05:33 PM, Xin Tong wrote:
>> On Tue, Feb 7, 2012 at 12:09 PM, Andrew Haley <aph@redhat.com> wrote:
>>> On 02/07/2012 04:56 PM, Xin Tong wrote:
>>>> I am wondering how gcc accesses global variables on x86. from the code
>>>> i have seen so far, it seems to use the %RIP as the base register. Is
>>>> it always like this?
>>>>
>>>>   4006f9:       f0 80 0d bf 0b 20 00    lock orb $0x0,0x200bbf(%rip)
>>>>      # 6012c0 <t2lockor>
>>>>   400700:       00
>>>>
>>>> t2lockor is a global variables.
>>>
>>> This is x86_64, I think.  The answer is that it depends on whether you
>>> are using PIC, and the model you're using.  Try -mcmodel=large for a
>>> variation.  PC-relative loads are convenient for everything except the
>>> large memory model.
>>
>> What are large memory model and PIC, can you please briefly explain.
>
> For large memory model see -mcmodel in the gcc docs.
> PIC means position independent code, as in the gcc command -fpic.
>
>> the PC of instructions are going to change in the linkage stage. so
>> the linker patches the offset if rip is used to access the global
>> variables.
>
> Yes.
>
> Andrew.